Variable Length Motif-Based Time Series Classification

Abstract

Variable length time series motif discovery has attracted great interest in the community of time series data mining due to its importance in many applications such as medicine, motion analysis and robotics studies. In this work, a simple yet efficient suffix array based variable length motif discovery is proposed using a symbolic representation of time. As motif discovery is performed in discrete, low-dimensional representation, the number of motifs discovered and their frequencies are partially influenced by the number of symbols used to represent the motifs. We experimented with 4 electrocardiogram data sets from a benchmark repository to investigate the effect of alphabet size on the quantity and the quality of motifs from the time series classification perspective. The finding indicates that our approach can find variable length motifs and the discovered motifs can be used in classification of data where frequent patterns are inherently known to exist. © Springer International Publishing Switzerland 2014.